Fashion Hash Net

Preamble

I completed my graduation project based on this responsitory.

Here are the details of my work. And this is the link to the original project.

Description

This responsitory contains the code of paper Learning Binary Code for Personalized Fashion Recommendation

Required Packages

• pytorch
• torchvision
• PIL
• numpy
• pandas
• tqdm: A Fast, Extensible Progress Bar for Python and CLI
• lmdb: A universal Python binding for the LMDB 'Lightning' Database.
• yaml: PyYAML is a full-featured YAML framework for the Python programming language.
• visdom: To start a visdom server run python -m visdom.server

I upgraded the version of PyTorch to 1.2.0 and the package dependency is solved automatically with conda.

The last 4 packages can be install via conda:

conda install python-lmdb pyyaml visdom tqdm -c conda-forge

conda create -n FashionHashNet python>=3.8 anaconda
conda activate FashionHashNet

pip install -r requirements.txt
conda install --file requirements.txt

Remember to add this project in the PYTHONPATH environmental variable if you plan to run the experiments on the terminal:

export PYTHONPATH=$PYTHONPATH:/path/to/project/folder

Docker

docker build -t fashion-hash-net:v1 .
docker run -itd --shm-size=256M --name fashion-hash-net fashion-hash-net:v1

If you're trying to train networks with this image, you may run into an error like:

ERROR: Unexpected bus error encountered in worker. This might be caused by insufficient shared memory (shm).

This is because Docker defaults to 64MB of SHM no matter how much memory it's allocated. You can override this by passing e.g. --shm-size 32g as an argument to docker run before the Docker image name.

How to Use the Code

The main script scripts/run.py currently supports the following functions:

ACTION_FUNS = {
    # train models
    "train": train,
    # runing the FITB task
    "fitb": fitb,
    # evaluate pairs accuracy
    "evaluate-accuracy": evalute_accuracy,
    # evaluate NDCG and AUC
    "evaluate-rank": evalute_rank,
    # compute the binary codes
    "extract-features": extract_features,
}

Configurations

There are three main modules in polyvore:

• polyvore.data: module for polyvore-dataset
• polyvore.model: module for fashion hash net
• polyvore.solver: module for training

For configurations, see polyvore.param, and we give some examples in cfg folder. The configuration file was written in yaml format.

Train

To train FHN-T3 with both visual and semantic features, run the following script:

start the visdom server before starting the training

python -m visdom.server > visdom.log 2>&1 &

scripts/run.py train --cfg ./cfg/train/FHN_VSE_T3_630.yaml

python scripts/run.py train --cfg ./cfg/train/FHN_T3_53.yaml > train_53.log 2>&1 &

Evaluate

To evaluate the accuracy of positive-negative pairs:

scripts/run.py evaluate-accuracy --cfg ./cfg/evaluate-accuracy/FHN_VSE_T3_630.yaml

python scripts/run.py evaluate-accuracy --cfg ./cfg/evaluate-accuracy/FHN_T3_53.yaml > evaluate_accuracy_53.log 2>&1 &

To evaluate the rank quality:

scripts/run.py evaluate-rank --cfg ./cfg/evaluate-rank/FHN_VSE_T3_630.yaml

python scripts/run.py evaluate-rank --cfg ./cfg/evaluate-rank/FHN_T3_53.yaml > evaluate_rank_53.log 2>&1 &

To evaluate the FITB task:

scripts/run.py fitb --cfg ./cfg/fitb/FHN_VSE_T3_630.yaml

python scripts/run.py fitb --cfg ./cfg/fitb/FHN_T3_53.yaml > fitb_53.log 2>&1 &

How to Use the Polyvore-$U$s

1. Download the data from OneDrive and put the polyvore folder under data;

2. Unzip the polyvore/images/291x291.tar.gz;

3. Use script/build_polyvore.py to convert images and save in data/polyvore/lmdb.

python scripts/build_polyvore.py data/polyvore/images/291x291 data/polyvore/images/lmdb

The lmdb format can accelerate the load of images and set as default in configuration. If you don't want to use the lmdb format, change the setting to use_lmdb: false in yaml files.

See <data/README.md> for details

How to Cite

@inproceedings{Lu:2019tk,
author = {Lu, Zhi and Hu, Yang and Jiang, Yunchao and Chen, Yan and Zeng, Bing},
title = {{Learning Binary Code for Personalized Fashion Recommendation}},
booktitle = {CVPR},
year = {2019}
}

Contact

Email: [email protected]